Complex Data Input Ordering for Neural Image Reconstruction
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing medical imaging technologies face challenges in reconstructing high-quality images when data acquisition does not meet the Nyquist criterion, leading to artifacts such as aliasing and streaking, particularly in modalities like MRI and CT, where under-sampling is inevitable or intentional to reduce scan time.
Innovation Solution
A data-driven manifold learning approach, referred to as AUTOMAP, transforms raw sensor data into medical images without constraining the reconstruction process to human-devised, acquisition-specific transforms, using neural networks to reorder and process data from a sensor domain to an image domain.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Productivity
If under-sampling is used to reduce scan time, then productivity is improved, but measurement precision deteriorates due to artifacts
Solution Approach 1:
A neural network is introduced as an intermediary between the undersampled k-space data and the final image reconstruction. The neural network learns to map undersampled data to high-quality images, effectively mediating the transformation while compensating for the deficiencies of undersampling. This allows rapid scanning with reduced artifacts while maintaining diagnostic image quality.
2Manufacturing precision
If conventional reconstruction methods are used, then manufacturing precision is maintained, but adaptability deteriorates due to modality-specific constraints
Solution Approach 1:
The neural network framework is designed to be universal across multiple imaging modalities (MRI, CT, optical imaging). By training the network on diverse modalities and data types, a single system can handle different imaging techniques without requiring modality-specific reconstruction algorithms, thereby achieving both high precision and broad adaptability.
Solution Approach 2:
The system adapts to different imaging modalities by changing key parameters such as input data format, network architecture configuration, and training parameters. This allows the same fundamental neural network approach to be applied across MRI, CT, and optical imaging by simply adjusting these parameters rather than developing separate reconstruction methods for each modality.
3Measurement precision
If data is reordered with real and imaginary components adjacent, then neural network processing effectiveness is improved, but device complexity increases
Solution Approach 1:
The real and imaginary components of the complex data are reordered to be adjacent to each other before being fed into the neural network. This preliminary reorganization of data structure simplifies the neural network's processing by ensuring that related data elements are positioned contiguously in the input tensor, improving computational efficiency and reconstruction accuracy without adding significant system complexity.
Data Source
AI summary
Systems, methods, and media for complex input data configurations for imaging applications. Complex data optimization can be provided to improve accuracy of models (e.g., neural networks) used to reconstruct medical images from raw sensor data, for example. Complex data optimization can include applying raw sensor data to an input layer of a neural network to generate an input vector ordered such that real components and imaginary components of samples in the raw sensor data are adjacent. The input vector can then be applied to convolutional layer of the neural network.